Nature Cancer's DeepFAN: AI nodule reading aids judged by how they change radiologists
jiqizhixin · x · 2026-11-07
- DeepFAN, a Nature Cancer study involving Peking Union Medical College Hospital, HKU, Peking University People's Hospital, Deepwise and collaborators, argues diagnostic models should be evaluated by how they change readers' decisions, not just standalone scores.
- Architecture: a vision transformer captures nodule context, a 3D convolutional branch learns local imaging attributes, and a graph network relates features to predict malignancy probability.
- Paired assistance trial: 12 radiologists (1–5 years experience) read 463 pathology-confirmed nodules from 400 patients across three centers, with and without AI, in two crossover rounds with a four-week washout.
- Model vs reader outcomes: standalone AUC 0.9+, reported separately from reader performance.
More from Research
- SAIGE-QTL, a mixed-model tool for single-cell eQTL mapping, published in Nature Genetics — anshulkundaje · 2026-11-08
- Fleuret and giffmana debate numerical blow-up risks in an activation-norm formula — giffmana · 2026-11-08
- MaxxBench: models clustered at 74% on DeepSWE spread 23 points when tasks change slightly — dyn___ · 2026-11-08
- Two ~500B MoE Post-Training Designs Compared: Separate RL Runs vs Joint RL — himanshustwts · 2026-11-08
- Stories about a generic AI beat self-referential stories, challenging the Persona Selection Model — repligate · 2026-11-08
- Reviewer refuses to read 70-page appendices: answer key questions in the main paper — CharlotteHase · 2026-11-07